You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: lib/shape_inference.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -167,7 +167,7 @@ Simplification of an inequality, and constraint propagation, can generate more c
167
167
* Stage 1 is online as tensors are composed, and conservatively performs unification and constraint propagation. Stages 2, 3, 4 are only performed once necessary: when projections or dimensions are requested.
168
168
* Stage 2, when solving the constraints, substitutes dim variables in terminal shapes that do not have a LUB or other constraints, by dimension-1. (This is generalized at stage 6 to all variables.) (FIXME: reconsider this, see the algo for row variables: a new LUB can still be inferred.) Forces coefficients coming from precision byte sizes.
169
169
* Stage 3, when solving the constraints, sets yet-unknown dimension and row variables in terminal shapes to their least upper bounds (if any), but for rows only if they don't have a `Total_elems 1` constraint. It substitutes row variables in terminal shapes that do not have a LUB by one axis if that's required to satisfy the variable's constraint.
170
-
* Stage 4 sets yet-unknown dimensions with >1 lower bounds from direct accesses, to their LUBs if they have any. It substitutes row variables in terminal shapes that do not have a LUB by no-further-axes. (This is generalized at stage 6 to all variables.) At this stage, we inject `Shape_row` constraints into the inequalities, so that we can re-process the variables of interest without traversing the whole environment.
170
+
* Stage 4 sets yet-unknown dimensions with >1 lower bounds from direct accesses, to their LUBs if they have any. It substitutes row variables in terminal shapes that do not have a LUB by no-further-axes. (This is generalized at stage 6 to all variables.) In Total_elems constraints with multiple row variables, it substitutes row variables originating from axes of non-output kind, and which do not have a LUB, by no-further-axes -- otherwise these constraints can be too hard to unlock. At this stage, we inject `Shape_row` constraints into the inequalities, so that we can re-process the variables of interest without traversing the whole environment.
171
171
* Stage 5 addresses `Total_elems` and `Exact` constraints with yet-unknown row variables. For `Total_elems` and a single row variable: if the constraint can be satisfied by assuming the row variable is no-further-axes, it sets the row variable to `Broadcastable`, otherwise it sets it to one axis of the required dimension. For multiple row variables, if one is of the Output kind, sets the other variables to no-further-axes, and retries.
172
172
* Stage 6 sets row variables in the remaining inequalities and updated shapes to no-further-axes values. This can unlock further between-axis inequalities because of row variables sandwiched between leftmost axes from their side of the inequality and rightmost axes from the other side of the inequality. In row constraints, this also unlocks inference for the embedded dim variables.
173
173
* Stage 7 sets all dim variables remaining in updated shapes to the lower bound if they have any, otherwise to dimension-1.
0 commit comments